Multilinguals at SemEval-2022 Task 11: Transformer Based Architecture for Complex NER

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 1623 - 1629 July 14-15, 2022 ©2022 Association for Computational Linguistics Multilinguals at SemEval-2022 Task 11: Transformer Based Architecture for Complex NER Amit Pandey*, Swayatta Daw*, and Vikram Pudi International Institute of Information Technology, Hyderabad, India {amit.pandey, swayatta.daw}@research.iiit.ac.in vikram@iiit.ac.in Abstract We investigate the task of complex NER for the English language. The task is non-trivial due to the semantic ambiguity of the textual structure and the rarity of occurrence of such entities in the prevalent literature. Using pre- trained language models such as BERT, we obtain a competitive performance on this task. We qualitatively analyze the performance of multiple architectures for this task. All our models are able to outperform the baseline by a significant margin. Our best performing model beats the baseline F1-score by over 9%. 1 Introduction The Named Entity Recognition (NER) task aims to detect entities from unstructured text and clas- sify them into predefined categories. Although the task of NER has been investigated adequately by previous research work (Mansouri et al., 2008; Nadeau and Sekine, 2007; Lample et al., 2016a; Florian et al., 2003; Ritter et al., 2011), the de- tection of named entities in open-domain settings is non-trivial. Moreover, the introduction of addi- tional layers of complexity, in the form of semantic ambiguity and a lower amount of contextual avail- ability, poses further challenges. For example, in a low-context and semantically ambiguous sentence such as Let us play Among Us, the token sequence Among Us, can refer to a common phrase or a pop- ular video game, and hence be categorized as a Creative Work (CW). Recently, deep learning models have gained pop- ularity for NER (Yadav and Bethard, 2019; Li et al., 2020; Habibi et al., 2017). However, these ap- proaches are data-intensive and become ineffective when there is a lack of labeled data. To foster re- search in this area, (Malmasi et al., 2022b) has introduced the SemEval MultiCoNER shared task that deals with multilingual complex named en- *Equal Contribution tity recognition. This task in based on the com- plex NER, search query and code-mixing NER challenges introduced by Meng et al. (2021) and Fetahu et al. (2021). The baseline introduced by the organizers for this MultiCoNER challenge is a pre-trained XLM-RoBERTa model (Conneau et al., 2019) that was further fine-tuned on the task- specific training dataset. This paper describes our approach to tackle com- plex NER task for the English language using state- of-the-art deep learning models and introduces a simple neural network architecture that builds on top of pre-trained language models. We compare multiple architectures on the validation and test set of the shared task. All our models outperform the baseline by a significant margin. Through our ex- periments, we discover that leveraging transformer models based on attention mechanism (Vaswani et al., 2017) results in better performance even in low context and ambiguous settings. The code is available at https://github.com/ AmitPandey-Research/Complex_NER We describe the prior research work done with respect …

Content truncated to 500 words